[2501.13986] An Efficient Sparse Kernel Generator for O(3)-Equivariant Deep Networks

PDF view of the paper entitled Kernel Mawlid effective scattered for deep networks O (3)

PDF HTML (experimental) view

a summary:Neurological networks of rotation, IE networks designed to ensure some engineering relations between their inputs and outputs, are returning to technical performance on the tasks of deep spatial learning. It shows high data efficiency during training and reduces the time to significantly to the potential calculations between rotation compared to classic methods. The key to these models is the Clebsch-Gordon (CG) product, which is Kernel, which shrinks two dense feature vehicles with a very full-time tensioner organized to produce a thick production transmission. The process, which may repeat millions of times for typical model models, is the neck of the inexpensive and ineffective bottle. We offer the GPU scattered generator for the CG Tensor product that provides large speeds on the best open and closed applications. Our implementation achieves high performance by managing a carefully limited GPU memory through fixed analysis at the time of typical translation, which reduces readings and writing to global memory. We divide the tensioner product into a series of smaller nucleus with transactions that are fully suitable for records, allowing us to emit long account instructions flows that increase parallel at the level of instructions. By integrating the CG Tensor product with the subsequent graphic damage, we reduce both the medieval storage and the global memory movement on the naive methods that repeat the input data. We also offer an improved nucleus to include the CG Tensor product and a new identity for the upper partial derivatives needed to predict the joint powers between the growth. Our nucleus provides up to 1.3X acceleration on the closed NVIDIA closed source package, as well as accelerating 10x on the widely used E3nn package. In the FP64 resolution, we offer up to 6.2X accelerate the time of inference to the MACE Chemistry Foundation model on the original unlimited version.